This repository provides the official GGUF quantizations for JonusNattapong/LFM3-Omni-2B-abliterated, an omnimodal 2B parameter Hybrid (Conv + Attention) architecture abliterated to remove refusal steering vectors (100% uncensored).
Now featuring v10: "Panu" (ภาณุ) - Sovereign Multi-Dialect Thai AI Assistant with Long-Term Memory, Deep Reasoning (<think>), and Full-Duplex Voice Barge-in support.
Trained on 10,000 multi-dialect conversations covering Northern (Kam Mueang), Isan, Southern (Pak Tai), Korat (Tai Berng), Patani Malay (Yawi), Deep Thai R1 Reasoning, and OS Agent Tools.
| File Name | Quantization | Size | RAM / VRAM | Recommended Use |
|---|---|---|---|---|
LFM3-Omni-2B-v10-Sovereign-Q8_0.gguf |
Q8_0 | ~1.84 GB | ~2.6 GB | Highest Quality (99.9% FP16 parity) for desktops & servers |
LFM3-Omni-2B-v10-Sovereign-Q6_K.gguf |
Q6_K ⭐ | ~1.42 GB | ~2.0 GB | 6-bit High Fidelity: Optimal balance of quality and size |
LFM3-Omni-2B-v10-Sovereign-Q5_K_M.gguf |
Q5_K_M | ~1.25 GB | ~1.8 GB | Low-memory balanced quantization |
LFM3-Omni-2B-v10-Sovereign-Q4_K_M.gguf |
Q4_K_M | ~1.09 GB | ~1.5 GB | Fastest & Lightest; ideal for edge devices, laptops, mobile |
Required when running vision or live speech inference via llama-server / llama-cli:
| File Name | Type | Size | Description |
|---|---|---|---|
mmproj-LFM3-Omni-2B-Q4_K_M.gguf |
mmproj | ~278 MB | Vision & Audio projector for multimodal inference in llama.cpp |
| File Name | Quantization | Size | RAM / VRAM | Recommended Use |
|---|---|---|---|---|
LFM3-Omni-2B-abliterated-v6-Q8_0.gguf |
Q8_0 | ~1.76 GB | ~2.5 GB | v6 Baseline text release |
LFM3-Omni-2B-abliterated-v6-Q4_K_M.gguf |
Q4_K_M | ~1.05 GB | ~1.6 GB | v6 Baseline lightweight release |
llama.cppllama-cli \
-m LFM3-Omni-2B-v10-Sovereign-Q6_K.gguf \
--mmproj mmproj-LFM3-Omni-2B-Q4_K_M.gguf \
-p "<|im_start|>user\nสวัสดีครับภาณุ แนะนำอาหารเหนือลำๆ ให้หน่อยครับ<|im_end|>\n<|im_start|>assistant\n" \
-n 512
llama-server \
--model LFM3-Omni-2B-v10-Sovereign-Q6_K.gguf \
--mmproj mmproj-LFM3-Omni-2B-Q4_K_M.gguf \
--alias lfm3-omni \
--host 127.0.0.1 \
--port 18084 \
--ctx-size 8192 \
--jinja
JonusNattapong/LFM3-Omni-2B-abliterated